PALMPRINT VERIFICATION USING CONSISTENT ORIENTATION CODING
Zhenhua Guo1, Wangmeng Zuo2, Lei Zhang1, David Zhang1
1
Biometrics Research Centre, Department of Computing, the Hong Kong Polytechnic University
2
School of Computer Science and Technology, Harbin Institute of Technology
ABSTRACT
Developing accurate and robust palmprint verification
algorithms is one of the key issues in automatic palmprint
recognition systems. Recently, orientation based coding
algorithms, such as Competitive Code (CompCode) and
Orthogonal Line Ordinal Features (OLOF), have been
proposed and have been attracting much research attention.
Such algorithms could achieve high accuracy with high
feature matching speed for real time implementation. By
investigating the relationship between these two different
coding schemes, we propose in this paper a feature-level
fusion scheme for palmprint verification. Only the stable
features which are consistent between the two codes are
extracted for matching. The experimental results on the
public palmprint database show that the proposed fusion
code could achieve at least 14% EER (Equal Error Rate)
reduction compared with either of the original codes.
personal identification. Fig. 1-a shows the Region of
Interest (ROI) of a palmprint sample image [2].
Competitive Code (CompCode) [8] and Orthogonal Line
Ordinal Features (OLOF) [9] are state-of-the-art algorithms
for palmprint verification because they could achieve very
high accuracy with a high matching speed. Taking palm
lines as negative lines, Kong and Zhang [8] proposed to
extract and code the orientation of palm lines for palmprint
verification. They applied six Gabor filters to the image and
selected one main orientation for each local region. Sun et
al. [9] proposed a new palmprint representation. They
compared two elongated line-like image regions, which are
orthogonal in orientation, and then assigned one bit feature
code. Three bits are then obtained by using three different
orientated ordinal filters. Two example maps of CompCode
and OLOF extracted from Fig. 1-a are shown in Fig. 1-b
and Fig. 1-c.
Index Terms—palmprint verification, feature-level
fusion, orientation coding
1. INTRODUCTION
As a substitution or complementary technology for
traditional personal authentication methods, biometric
techniques are becoming more and more popular in public
security applications. Biometrics is the study of methods for
uniquely recognizing humans based upon one or more
intrinsic physical or behavioral traits, including the
extensively studied fingerprint, face, iris, speech, hand
geometry, etc [1]. Palmprint authentication as an emerging
biometrics technology has been drawing much attention in
both academic and industry societies [2-10] because it owns
many merits, such as robustness, user-friendliness and costeffectiveness.
Palmprint is composed of three main kinds of features:
principal lines (usually three dominant lines on the palm),
wrinkles (weaker and more irregular lines) and minutia
(ridge and valley features which are similar to those in
fingerprint images) [2]. Unlike fingerprint which requires a
high resolution (about 500dpi) to capture clear minutia
features, the principal lines and wrinkles in a palmprint
could be captured under a low resolution (<100dpi) and
they could provide enough discriminatory information for
(a)
(b)
(c)
Fig.1 (a) The ROI of a sample palmprint image and the
extracted (b) CompCode map and (c) OLOF map. Different
features are represented by using different gray values.
From Fig. 1, we can see that the OLOF and CompCode
maps of the same palmprint are different but both of them
have clear representation of principal lines and wrinkles.
Since principal lines and wrinkles are the most stable and
robust features in low resolution palmprint images, it
inspires us to develop a feature level fusion scheme to
further enhance those stable features for a more robust
feature extraction. After investigating the correlation
between OLOF and CompCode, in this paper, we propose
such a feature level fusion scheme according to the
consistency between the two different codes.
The rest of the paper is organized as follows. Section 2
briefly reviews CompCode and OLOF. Section 3 analyzes
the correlation between them and then fuses them. Section 4
verifies the proposed scheme on a large public palmprint
database. Finally, Section 5 concludes the paper.
2. BRIEF REVIEW OF COMPCODE AND OLOF
each pixel and a three-bit ordinal code is obtained according
to the sign of the filtering results. The matching metric on
Hamming distance is defined the same as Eq. 3.
2.1. CompCode [8]
To extract the orientation information from palm lines,
CompCode uses six real parts of the neurophysiology-based
Gabor filters ψ θ , which is defined as:
ω2
− 2 (4 x′
ω
ψ ( x, y, ω ,θ ) =
e 8κ
2πκ
2
+ y ′2 )
κ2
⎛
−
⎜ eiω x′ − e 2
⎜
⎝
⎞
⎟
⎟
⎠
(1)
,
where
x′ = ( x − x0 ) cos θ + ( y − y0 ) sin θ
y ′ = −( x − x0 ) sin θ + ( y − y0 ) cos θ . (x0,y0) is the center of
the function; ω is the radial frequency in radians per unit
length and θ is the orientation of the Gabor function in
⎛ 2δ + 1 ⎞
radians. κ is defined by κ = 2 ln 2 ⎜⎜ δ
⎟⎟ , where δ is
⎝ 2 −1 ⎠
the half-amplitude bandwidth of the frequency response.
For each pixel, six filters with θ i = jπ / 6 ,
j={0,1,2,3,4,5}, are applied. According to palm lines’
property, CompCode selects:
I CompCode = arg min( I ( x, y ) *ψ R ( x, y, ω , θ j ))
j
(2)
as the orientation at position (x, y) of image I. To implement
real-time palmprint recognition, CompCode uses three bits
to represent each orientation [8]. An angular distance based
on Hamming distance is used for distance comparison:
M
N
3
∑∑∑ ( P ( x, y) ⊗ Q ( x, y))
i
D ( P, Q ) =
b
b
i
y =1 x =1 i =1
(3)
3* M * N
where P and Q are two CompCodes. Pib (Qib ) is the ith bit
plane of P(Q) and ⊗ is bitwise exclusive OR.
2.2. OLOF [9]
OLOF uses an orthogonal line ordinal filter to compare two
orthogonal line-like palmprint image regions. The filter is as
follows:
π
OF (θ ) = f ( x, y, θ ) − f ( x, y, θ + )
2
f ( x, y , θ ) =
2
⎛ x cosθ + y sin θ ⎞ ⎛ x cosθ + y sin θ
−⎜
⎟ −⎜⎜
σ
σy
x
⎠ ⎝
e ⎝
⎞
⎟
⎟
⎠
2
(4)
where θ denotes the orientation of the 2D Gaussian filter,
σ x and σ y are the horizontal and vertical scales of the
filter.
Similar to CompCode, three ordinal filters OF (0) ,
OF (π / 6) and OF (π / 3) (refer to Fig. 2) are applied to
Fig.2 Ordinal filters.
An integer value could be computed with the three-bit
code using the following formula:
I OLOF = s (W (0)) *1 + s (W (π / 6)) * 2 + s (W (π / 3)) * 4
⎧1, x ≥ 0
s ( x) = ⎨
⎩0, x < 0
(5)
where W (θ ), (θ = 0, π / 6, π / 3) is the filter response after
convolving OF (θ ) with the input image.
3. CONSISTENT ORIENTATION CODE
CompCode and OLOF could achieve high accuracy.
Although they use different coding strategies, they have
some consistency in representing palm stable features, as
can be seen in Fig. 1. After investigating the correlation
between these two codes, we propose a feature-level fusion
scheme, namely consistent orientation code, in this section.
3.1. Correlation between CompCode and OLOF
Although OLOF is not designed as an orientation estimator,
it can be used to represent a line’s orientation like the
CompCode does. Suppose there is a straight dark line in a
white background as shown in Fig. 3-a. We rotate Fig. 3-a
counterclockwise by different angles, e.g. from 1 to 180
degree with 1 degree interval (referring to Fig. 3-b ~ Fig. 3f for examples). The associated I CompCode and I OLOF code
maps are plotted in Fig. 4. It can be seen that there is clear
correlation between CompCode and OLOF for this simple
line image. Taking Fig. 3-a for example, since the dark line
is closer to the “+” part than the “-” part of OF (0) and
OF (π / 6) (referring to Fig. 2), the filtering outputs of
OF (0) and OF (π / 6) will be smaller than 0, while the
filtering output by OF (π / 3) will be bigger than 0. So
according to Eq. 5, I OLOF is equal to 4. Meanwhile, the
competitive code, I CompCode , of this line will be 3 because
ψ R ( x, y, ω ,θπ / 2 ) has the minimal value.
For real palmprint image, however, the structure is
much more complex. For example, in a local region, there
may have non-straight lines, weak lines and even multiple
lines. Table 1 shows the co-occurrence percentage between
CompCode and OLOF on a portion of the PolyU database
[11] (first six images of each palm, totally 386*6=2316
images). As we can see from Table 1, there is a strong
correlation between CompCode and OLOF but it is hard to
find a direct mapping function between them. Intuitively,
the consistent parts between the two codes will reflect the
most stable and important features in the palmprint, while
the inconsistent parts may come from noise and the unstable
features in the palmprint. Therefore, it inspires us to
consider a feature-level fusion scheme to extract the
consistent features in the two codes and remove the nonconsistent part so that the recognition accuracy could be
further improved.
consistent part between CompCode and OLOF extracted
from Fig. 1. As we can see, the strong consistent part is
mainly from principal lines and strong wrinkles.
However, the strong consistency is less than 50 percent.
If we keep this part only, we may not have enough
discriminate information for verification. On the other hand,
if we regard that the part highlighted with bold number in
Table 1 has a relatively weak consistency, the valid portion
is more than 90 percents (93.5562%, the sum of bold
number). The weak consistency can contribute additional
information while some unstable parts are removed. Fig. 5-b)
shows the weak consistent part. We can see that weak
consistency contains almost all of principal lines and
wrinkles.
(a)
(b)
(c)
(d)
(e)
(f)
Fig. 3 An image with a straight line and its rotated image. (a) is
the original image and (b)~(f) are the 30, 60, 90, 120 and 150
degree counterclockwise rotated versions of (a).
8
OLOF Value
7
CompCode Value
(a)
(b)
Fig. 5 (a) Strong consistent part (in black color) between Fig. 1b and Fig. 1-c; and (b) weak consistent part (including strong
consistent part) between Fig. 1-b and Fig. 1-c.
6
Code Value
5
4
Here, we propose two different fusion schemes: Strong
Consistent Fusion (keep strong consistent part only) and
Weak Consistent Fusion (keep weak consistent part only).
The distance for the proposed fusion scheme is defined as:
3
2
M
D ( P, Q ) =
0
0
20
40
60
80
100
Rotation Angle
120
140
160
180
3
Table 1. Co-occurrence percentage of CompCode and OLOF
on partial PolyU palmprint database.
CompCode
0
1
2
3
4
5
OLOF
0.4220
3.7728
0.0055
7.2229
0.1733
0.0388
0.4398
3.7095
0.1733
0.4797
0.0376
4.2097
0.3500
0.0063
3.8072
8.4643
0.4559
0.1955
0.0056
0.4286
3.6424
0.0401
8.2218
3.8915
3.7009
0.4171
0.0391
0.1922
7.4978
0.0062
3.6534
0.4063
b
b
M ( x, y ) ∩ QM ( x, y )) ∩ ( Pi ( x, y ) ⊗ Qi ( x, y ))
y = 0 x = 0 i =1
M
(6)
N
∑∑ P
M ( x, y ) ∩ QM ( x, y )
3
y =0 x =0
Fig. 4 Rotation angle vs. Coding value of OLOF and
CompCode.
0
1
2
3
4
5
6
7
N
∑∑∑ (P
1
8.0190
3.9493
0.0063
0.4390
3.5182
0.0393
0.4342
0.1823
3.7362
8.4655
0.0377
4.0738
0.3580
0.0070
0.1822
0.4450
3.2. Proposed Fusion Scheme
Table 1 shows that around half (47.8912%, the sum of
underscored number) of the CompCode and OLOF codes
has a strong consistency. Fig. 5-a) shows the strong
where I is bitwise AND operation, PM and QM are the
corresponding masks to indicate whether this pixel is kept
according to the consistency criteria.
4. EXPERIMENT RESULTS
To validate our proposed consistent orientation coding
scheme, we tested it on a large public palmprint database
[11]. The database contains 7752 grayscale images collected
from 386 palms. Those palmprint images were collected in
two sessions with an average time interval of 69 days.
Around 10 images per palm were captured per session.
The central part (128*128) of the palm image is
cropped by using an ROI extraction algorithm similar to [2].
To further reduce the influence of imperfect ROI extraction,
we translate the feature map vertically and horizontally from
-4 to 4 when matching it with the feature maps in the
database. The minimal distance obtained by translated
matching is regarded as the final distance. Compared with
originally used searching range, -2 to 2 [8-9], here the
searching range is increased for better accuracy.
In palmprint verification test, each palmprint image is
matched with all the other palmprint images in the database.
A match is counted as correct if the two palmprint images
are from the same palm; otherwise, the match is counted as
incorrect.
The
total
number
of
matches
is
7752×7751/2=30,042,876, and among them there are
74,068 correct matchings. The index of equal error rate
(EER), a point when false accept rate (FAR) is equal to
false reject rate (FRR), is used to evaluate the performance.
In addition, the discriminating index d ' (d-prime) [12] is
computed for reference.
The curves of receiver operating characteristic (ROC)
for the original CompCode, OLOF and the proposed two
fusion codes are plotted in Fig. 6. The indices of EER and
d ' are listed in Table 2. Since we increased the matching
range and made some optimization on ROI extraction, such
as some morphological operations after binarization to
overcome the broken finger problem caused by shading, the
result of CompCode is better than what was reported in the
original publication [8].
Table 2 Accuracy comparison.
Algorithm
EER
d'
CompCode
OLOF
Strong Consistent
Weak Consistent
0.0379
0.0553
0.0987
0.0324
5.4122
6.0301
7.0468
5.6925
FRR (when
FAR=3.3x10-6%)
1.2273
1.5621
4.4108
0.7858
4.5
4
3.5
FRR(%)
3
CompCode
OLOF
Weak Consistent Fusion
Stong Consistent Fusion
2.5
2
1.5
1
0.5
0
-6
10
-4
10
-2
10
FAR(%)
0
10
2
10
Fig. 6 Verification performance comparison.
From the experimental results, we can see that the weak
consistent fusion scheme could achieve much better
performance than the original CompCode and OLOF codes,
which is largely attributed to the removal of unstable or
non-robust information. Although strong consistent fusion
could get the highest d ' , it also has the highest EER value.
It validates that d ' is not highly correlated with ROC curve.
The proposed method achieves at least 14% (0.0379%-
>0.0324%) EER reduction and correctly reject more than
1,600 (0.0055%*29,968,808) impostor attempts compared
with either of the original codes.
5. CONCLUSION
In this paper, we analyzed the correlation between two
state-of-the-art
palmprint
verification
algorithms:
CompCode and OLOF. Based on the observations, a feature
level fusion scheme by exploiting the consistency between
the two codes was proposed. The experimental results on
the public database showed that by keeping the consistent
orientation features, the palmprint verification accuracy
could be improved significantly compared with either the
original CompCode or OLOF algorithm.
6. ACKNOWLEDGEMENT
This work is partially supported by the Hong Kong RGC General
Research Fund (PolyU 5351/08E).
7. REFERENCES
[1] Jain, A., Bolle, R., and Pankanti, S. (eds.), Biometrics:
Personal Identification in Networked Society, Kluwer Academic
Publishers, Boston, 1999.
[2] Zhang, D., Kong, W. K., You, J., and Wong, M., “Online
palmprint identification,” IEEE Trans. Pattern Analysis and
Machine Intelligence, vol. 25, pp. 1041-1050, 2003.
[3] Ribaric, S., and Fratric, I., “A biometric identification system
based on eigenpalm and eigenfinger features,” IEEE Trans. on
Pattern Analysis and Machine Intelligence, vol. 27, pp. 1698-1709,
2005.
[4] Han, C.-C., Cheng, H.-L., Lin, C.-L., and Fan, K.-C., ”Personal
authentication using palm-print features,” Pattern Recognition, vol.
36, pp. 371-381, 2003.
[5] Connie, T., Andrew, T., and Goh, K., “An automated palmprint
recognition system,” Image and Vision Computing, vol. 23, pp.
501-505, 2005.
[6] Hu, D., Feng, G., and Zhou, Z., “Two-dimensional locality
preserving projections (2DLPP) with its application to palmprint
recognition,” Pattern Recognition, vol. 40, pp. 339-342, 2007.
[7] Wang, J., Yau, W., Suwandy, A., and Sung, E., “Person
recognition by fusing palmprint and palm vein images based on
“Laplacianpalm” representation,” Pattern Recognition, vol. 41, pp.
1531-1544, 2008.
[8] Kong, A., and Zhang, D., “Competitive coding scheme for
palmprint verification,” International Conference on Pattern
Recognition, pp. 520-523, 2004.
[9] Sun Z., Tan T., Wang Y., and Li. S. Z., “Ordinal palmprint
representation for personal identification,” IEEE Computer Society
Conference on Computer Vision and Pattern Recognition, pp. 279284, 2005.
[10] Jia, W., Huang, D.-S., and Zhang, D., “Palmprint verification
based on robust line orientation code,” Pattern Recognition, vol.
41, pp. 1521-1530, 2008.
[11]
PolyU
Palmprint
Database,
http://www.comp.polyu.edu.hk/~biometrics.
[12] Daugman J., and Williams G., “A proposed Standard for
biometric decidability,” Proc. CardTech/SecureTech Conference,
pp. 223-234, 1996.